Learning Reading Order via Document Layout with Layout2Pos

Anonymous

Abstract: Due to their remarkable performance, general-purpose multimodal pre-trained language models have gained widespread adoption for Document Understanding tasks. The majority of pre-trained language models rely on serialized text, extracted using either Optical Character Recognition (OCR) or PDF parsing. However, accurately determining the reading order of visually-rich documents (VrDs) is challenging, potentially affecting the accuracy of the extracted text and leading to sub-optimal performance in downstream tasks. For information extraction tasks, where entity recognition is commonly framed as a sequence-labeling task, incorrect reading order can hinder entity labeling. In this work, we avoid reading order issues by discarding sequential position information. Based on the intuition that layout contains the information for correct reading order, we present Layout2Pos–a shallow Transformer designed to generate position embeddings from layout. Incorporated into a BART architecture, our approach demonstrates competitiveness with models dependent on reading order across three benchmark datasets for information extraction. We also show that evaluating models using a reading order different from the one seen during training can result in substantial performance drops, thereby highlighting the importance of not relying on the reading order of documents.

Paper Type: long

Research Area: Information Extraction

Contribution Types: Publicly available software and/or pre-trained models

Languages Studied: English